Written by: Techub News Compilation

Introduction

At the Morgan Stanley 2026 TMT (Technology, Media, and Telecom) Conference, NVIDIA founder and CEO Jensen Huang participated in an in-depth interview lasting over 40 minutes. As a long-time supporter of the conference for 27 years, Huang's appearance comes after NVIDIA announced record quarterly earnings, having transformed its influence from a game graphics card manufacturer to the absolute core of global AI computing infrastructure. This conversation not only reviews NVIDIA's full-stack innovation journey over the past thirty-three years but also prospectively defines the new economic paradigm of the “AI factory,” detailing how computing power can directly translate into corporate revenue and global GDP, serving as a key window for understanding the next generation of AI-driven economy.

Summary

• NVIDIA's core competitiveness stems from the “full-stack” capability built over 33 years, integrating deep hardware, system architecture, and software libraries, allowing it to innovate the entire infrastructure annually.
• AI development has experienced three major turning points: generative, reasoning, and agent AI, with agent AI (like OpenClaw) leading to exponential growth in token consumption.
• The future data center is an “AI factory,” with tokens as the core output. The power limitations of the factory make the “tokens per watt” efficiency a critical metric for determining corporate revenue.
• The entire software industry will shift from tool leasing to “agent leasing,” creating a multi-trillion-dollar token economy, with ongoing and significant funding support for computing power demands.
• NVIDIA is expanding its frontier from digital AI to physical AI (robots, autonomous driving, digital biology), which will be the new frontier for the next decade.

From Game Graphics Cards to AI Supercomputers: Thirty-Three Years of Full-Stack Philosophy

The interview began with a reflection on NVIDIA's amazing growth scale. Morgan Stanley analysts noted that from the $48 million revenue at its IPO in 1998 to now hundreds of billions of revenue and net profit in a single quarter, this growth is “unprecedented.” Huang attributes all of this to the core principle established at the company's founding: creating a new computing platform.

He recalled that NVIDIA started with a specific algorithmic problem—computer graphics, specifically simulating light. This technology gave rise to the modern video game industry. “In a sense, we created the modern video game industry,” Huang stated. The early success was not coincidental but stemmed from a “full-stack” approach: NVIDIA not only designed chips (GPUs) but also created APIs that allowed applications to directly communicate with hardware (like the predecessor of DirectX) and even innovated system architecture (like AGP/PCIe bus) to enable 3D graphics acceleration on personal computers. This deep integration—tightly combining algorithms, chips, systems, software libraries, and developer tools—is in NVIDIA's DNA.

“Accelerated computing essentially requires full stack,” Huang emphasized. He referred to the practice of assigning engineers to game engines (like Epic's Unreal Engine) as “forward deployment engineers,” which ensured NVIDIA's technology was deeply integrated into final applications. This ability was later seamlessly migrated to the AI field. The birth of the world’s first AI supercomputer, the DGX-1, and the construction of large-scale clusters for Microsoft Azure and OpenAI are results of the same “full-stack, full-system” methodology.

Huang pointed out that it is this mastery of full stack that enables NVIDIA to innovate at an astonishing speed. “We don't just design a chip each year; we build a complete set of infrastructure.” From GPUs to CPUs (Grace), and to the connecting chip NVLink and networks (Spectrum-X AI Ethernet), NVIDIA owns and optimizes every layer of the stack. “If you don't own the whole stack, if you don't own other internal chips, it's hard to innovate every year. Because you're piecing together too many different things, and if you can't control it, it's a full-stack problem.”

The Three Turning Points of AI: From Generative, Reasoning to Agent Explosion

When asked about the dramatic changes in the AI field over the past two years, Huang clearly outlined three consecutive “turning points.” The first turning point is the popularization of generative AI, marked by the emergence of ChatGPT. It packaged powerful models (like GPT-3) through an easy-to-use interface and API, showcasing the capability of AI-generated content.

However, generative AI has the “hallucination” problem. Huang explained that this is not a fundamental flaw in technology but rather due to the generative process lacking context and factual basis. Thus, the second turning point followed: reasoning AI (represented by OpenAI's o1 model). This type of model introduced self-reflection and self-correction capabilities, and could perform “conditional generation” based on retrieved information, producing more reliable and factual answers. The immediate result was that the models became larger (about 10 times), the number of tokens generated skyrocketed (potentially hundreds of times), and the overall computing power demand surged by about a thousand times, while its usefulness led to a million-fold increase in usage.

Currently, we are at the third and possibly the most disruptive turning point: agentic AI. Huang cited OpenClaw as an example, noting that this open-source software surpassed the total download count of Linux within three weeks of its release, becoming the most popular open-source software in history, with a growth curve that is “almost vertical.” Its revolution lies in the user's instructions shifting from the past “what is” (query) to “what to do” (action).

“The current prompts are ‘create,’ ‘execute,’ ‘build,’ ‘write,’” Huang stated. Agent AI can understand complex task descriptions, autonomously conduct research, read manuals, use tools, and perform tasks. Within NVIDIA, they refer to these agents as “Claws.” These digital workers continuously run in the background, writing code, developing tools, consuming vast amounts of computing resources. “The number of tokens consumed by one agent could be millions of times more than the previous generative dialogues.” This indicates that corporate demand for computing power will no longer grow linearly, but will instead surge explosively.

AI Factory Economics: Computing Power Equals Revenue, Tokens Serve as Currency

Faced with such enormous demand for computing power, how to finance and build infrastructure becomes a key issue. Huang put forward his core assertion: the future data center is not a “data center,” but an “AI factory”.

“People don’t like building data centers because they don’t know what the return will be. But everyone loves building factories because factories make money.” This concept Huang proposed years ago has now become reality. The fundamental purpose of the AI factory is to produce “tokens,” which can be directly monetized. “We now know exactly that a company’s revenue is directly related to computing power. It's like how Mercedes-Benz is limited by factory capacity. If OpenAI had more computing power now, their revenue would be higher.” Thus, the first equation is: Computing Power = Revenue. From this, a larger equation can be derived: Computing Power = National GDP. No country would voluntarily give up the opportunity to possess intelligence.

In the factory model, the core constraint is power. The power supply for each factory is limited (like 1 gigawatt). Therefore, the most critical metric for measuring system performance is “tokens per watt.” Huang proudly claimed that NVIDIA's architecture is “an order of magnitude (10 times) ahead of alternatives” on this metric. This means that, under the same power budget, a factory using NVIDIA systems can produce 10 times the tokens, resulting in 10 times the potential revenue. Meanwhile, due to architectural efficiency, the “cost per dollar token” is also the lowest.

“For the first time in history, the choice of computer architecture in a company’s factory must be approved by the CEO,” Huang asserted. In an environment where power, land, and permits are all limited, choosing the wrong computing system will directly impact revenue for the following year. This scarcity forces customers to choose the most effective and reliable solutions, which is where NVIDIA has its advantage.

Trillion-Dollar Token Economy: Paradigm Shift in the Software Industry

So, where will the funding to support this vast AI factory construction come from? Huang painted a picture of fundamental transformation in the software industry.

He predicted that all future software will be “agentic”. Software companies will no longer simply rent out tools to customers (like CAD software licenses), but will lease “experts who can use these tools”—that is, AI agents. Just as modern enterprises have formal employees, contractors, and external experts, the future digital workforce will also consist of a mix of internally developed models (“in-house trained”), fine-tuned open-source models, and rented top-tier closed-source models (like OpenAI).

“Therefore, every software company—in the future—not only will be tool lessors but will also lease the agents that use these tools.” Huang cited examples like Cadence (electronic design automation), Synopsys (chip design), and Siemens (industrial software), pointing out that the business size of these companies will become “much larger,” but their business models will shift from software licensing to “token leasing.” The current IT/software industry, valued at trillions of dollars, consumes virtually no tokens, but in the future, it will become a “super consumer” of tokens.

“That’s where the funding comes from,” Huang concluded. The entire internet industry (cloud service providers) has shifted 100% of capital expenditure to generative/agent AI systems because they have been proven to improve the efficiency of all services, such as searching, shopping, advertising, and more. Now, the entire software industry will also join in, providing ongoing trillions of dollars of fuel for computing power demands. “We are at the starting point of this journey.”

Dealing with Constraints and Expanding Frontiers: Supply Chain, Investment, and Physical AI

In the face of industry-wide constraints such as chips, memory, power, and workforce, Huang showed an optimistic attitude. “I love constraints. Because in a constrained world, you have no choice but to choose the best.” Constraints force customers to make optimal decisions, which benefits NVIDIA, who possess the best architecture and complete delivery capability. He mentioned that NVIDIA's strong balance sheet has strategic significance, as the company has used capital to secure the entire supply chain from wafers, CoWoS packaging, memory to cables and capacitors in advance, enabling them to promise clients (like Microsoft’s Satya Nadella) quick deployment of gigawatt-scale AI factories.

Regarding ecosystem building, Huang emphasized that all of NVIDIA’s investments aim to expand and strengthen the CUDA ecosystem. He confirmed the agreement for NVIDIA to invest $30 billion in OpenAI (rather than the rumored $100 billion), and noted that OpenAI is expected to go public by the end of the year. Additionally, he mentioned ongoing efforts to expand capacity for OpenAI on AWS and the new demands arising from emerging AI labs (possibly referring to MSL).

Looking ahead, Huang believes that physical AI will be the new frontier for the next decade. NVIDIA is already leading in several physical AI fields such as autonomous driving (Alpamayo), robotics (Gr00t), digital biology (La-Proteina), and climate simulation (Earth-2). He predicts that in about two years, when agent AI becomes mainstream, the industry discussions will completely shift towards physical AI. “Building Lilly's (Eli Lilly) AI factory is impossible unless you possess NVIDIA’s full-stack software capabilities, model expertise, and domain knowledge.”

At the conclusion of the interview, when asked about his views on the company’s stock price, Huang reaffirmed his fundamental belief: “You cannot suppress it (the stock price). The reason is simple: computing power equals company revenue. In the future, every company will need computing power to generate income… computing power equals GDP.” He firmly believes that we are on the brink of a new economic era driven by computing power, executed by agents, and supported by a token economy, while NVIDIA's full-stack architecture is the cornerstone of it all.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。